If you embed a neural net trained to detect "some" kind of images, I would love to know how one could "dump" the list of things the net is looking for. Basically it's not how it works - the neural net is more or less a black box, you can't tell what kind of things it's looking for by just looking at its contents. You give it an image, it says it is a match or not, based on some internal algorithm - but looking at that algorithm tells you nothing about what kind of images would produce positive matches.
IMO, the politicians just want more authority and power. The GDPR gives them power to sue and fine large companies, ie, free money. I don't think it's about privacy any more than encryption backdoors are about fighting crime or photo scanning is about protecting children.
I think the Apple design used some fancy tricks to run the comparisons without ever decrypting the reference data and even if that wasn't possible almost every new device comes with some kind of "trusted" platform module that only "trusted" actors can access.
That in turn makes such a system far less useful to intelligence services and authoritarian regimes. You don't want to tell your target that you're looking for them and which keywords not to use!
I wonder if this is part one, and then a later revision will require centralised scanning rather than on-device scanning.